Enrich and qualify new inbound leads, then route them with a reason a person can check
When a lead fills in your form, an agent researches the company and the person, scores the lead against rules you wrote, and posts a short summary with its reasons to your sales channel or CRM. A person approves the route and any reply. Vercel and Anthropic describe running versions of this. Their warnings are that models do not always follow long rule lists, enrichment data is often stale, and quality drifts unless someone keeps sampling the output.
Each new lead gets a researched summary, a score, a route and a draft reply, and a person approves the action before anything changes in the CRM or reaches the lead.
- Difficulty
- Advanced
- Time
- A weekend to build, then weeks in shadow mode
- Sources
- 6 · 6 independent
- Updated
- 2026-10-13
The prompt
Fill the blanks here, or leave them for your agent to ask. Pick the tab for your agent. The hand-off button gives any agent this page as markdown.
You qualify inbound leads for [company], which sells [offer] to [buyer]. Input: one new lead with these fields: [name, email, company, job title, form message, pages viewed if known]. Do this: 1. Research. Read the company's website and up to two other public sources. Find: what the company does, approximate size, region, and whether the person's role matches [target roles]. Record the URL for each fact. If the email domain is personal (gmail.com and similar) or the company cannot be found, say so and do not guess. 2. Check the CRM. Look up [the domain] in [CRM]. If there is an open deal or a named owner, route to that owner and stop. 3. Score. Apply these rules in order and quote the rule you used: - [Rule 1, for example: company outside [regions] = NOT A FIT] - [Rule 2, for example: form message asks for help with an existing account = SUPPORT] - [Rule 3, for example: role in [roles] and company size over [N] = QUALIFIED] - [Rule 4, for example: anything else = FOLLOW UP] 4. Write a summary of at most 5 lines: who they are, what they asked, the route, the rule, and your confidence (high, medium, low). 5. Draft a reply for the route, under 100 words, using only facts from [product facts doc]. Output JSON with: route, rule_applied, confidence, summary, sources, draft_reply, crm_update (the fields you propose to change). Rules: - Do not write to the CRM and do not send anything. Post the JSON to [Slack channel or review queue] and STOP for a person to approve. - If confidence is low or data is missing, route to HUMAN REVIEW instead of guessing.
Human checkpoints: the agent stops and asks you
- Approve the qualification rules and routes before the agent scores any lead.
- Approve each CRM update and each reply in the first weeks (shadow mode).
- Review a sample of routed leads every week after shadow mode ends.
A wrong route can send a real buyer to a nurture list or a support case to sales. The agent writes to your CRM and may draft emails to real people. Keep CRM writes and all replies behind a human approval until you have reviewed enough output to trust it, and keep sampling after that.
What you need
- Written qualification rules and routesdata
Ask your best salesperson to write them. Vercel's first prompt came from its top SDR.
- CRM access with an API or connectoraccount
- A web research or enrichment sourcedata
The agent needs to read company websites at minimum. Enrichment data can be stale for small companies.
- A review channel such as Slack or a CRM queueapp
Steps
Write the rules with your best salesperson
List the routes (for example qualified, follow up, support, not a fit) and the rules for each one. Keep hard rules short and testable. Mark which rules depend on judgement.
Check: Every route has at least one rule, and every rule can be checked by reading a lead.
Put hard rules in code and judgement in the prompt
Region, company size thresholds and existing-owner lookups belong in code or a CRM lookup. Let the model handle only the parts that need reading, like what the form message asks for.
Check: The code step routes test leads correctly without the model.
Run in shadow mode
Let the agent research and score every new lead while your team keeps working leads the old way. Compare the agent's route with the human route each day. Vercel describes several weeks of this with its best SDR checking every output.
Check: You have a log of agent route against human route for each lead.
Fix the disagreements
Read every lead where the agent and the human differ. Add or sharpen a rule, or add the case to the prompt as an example. Watch for junior people at good companies, personal email domains, and wrong company size data.
Check: Disagreements drop and each one has a known cause.
Turn on approval mode
The agent now posts its summary, route and draft to your review channel. A person clicks approve before the CRM changes or the reply is sent.
Check: Nothing reaches the CRM or a lead without an approval click in the log.
Keep sampling
After you trust the output, review a sample of routed leads each week. Vercel reports moving from reviewing all output to sampling, and adding a second agent that watches for rule breaches.
Check: You have a weekly sample with a pass rate you write down.
What goes wrong
The model does not always follow the rules in a long prompt, so some leads get the wrong route.
Fix: Move hard rules into code or CRM lookups. Keep the prompt short. Vercel's COO said their prompt grew to about 1,000 lines before they moved logic into 14 code rules.
Read the lesson →Enrichment cannot resolve personal email domains or returns stale company size and funding data.
Fix: Route unresolved leads to human review. Show the source and date of each enrichment fact so the reviewer can judge it.
Scoring on headcount alone misroutes leads, for example a junior signup at a large company.
Fix: Include the person's role and the form message in the rules, not only firmographics.
Output quality drops slowly and nobody notices.
Fix: Keep a weekly sample review and log the pass rate. One vendor test names silent scoring drift as a failure mode, and SaaStr's write-up of Vercel warns that agents degrade quietly.
Read the lesson →An existing customer or open deal gets treated as a new lead.
Fix: Look up the domain in the CRM first and route to the current owner before any scoring.
Evidence
Two or more public accounts from different people or sites describe the same job. We summarise each source in our own words. Open the original before you rely on it.
- Blog2026-01-23· Custom agentBuilding Vercel's first GTM agent
A Vercel GTM engineer describes an agent that enriches inbound leads, does web research, sorts each lead into one of four buckets, posts its reasoning to Slack and drafts a reply for an SDR to review and send. He says the inbound SDR team went from 10 people to 1. He describes no agent failures but warns the CRM connection needs proper engineering.
Drew Bredvick · drew.tech - Blog2026-10-06· Custom agentHow to automate inbound
Tunguz reports Vercel COO Jeanne DeWitt Grosser's account. Humans reviewed all early outreach, then sampled. The prompt grew from about 125 to about 1,000 lines, the model did not always follow its rules, and the team moved qualification into 14 code rules with a second agent watching for breaches.
Tomasz Tunguz · tomtunguz.com - GitHub· Custom agentvercel-labs/lead-agent
An open-source reference build of a lead agent that researches a lead, classifies it as qualified, follow up or support with reasoning, and waits for a Slack approve or deny before sending. The README has no CRM integration, and some functions are placeholders. The repo is archived.
Vercel Labs · github.com - Blog2026-08-07· Claude CoworkHow Anthropic's business development team uses Claude to run inbound and outbound at scale
An Anthropic BDR describes a daily skill that researches new leads, scores them and drafts a first touch. Salesforce updates are proposed with evidence and wait for approval. He keeps a person on every send. The post gives no conversion numbers.
John Albert · claude.com - Blog2025-11How our AI BDR drove inbound from 30% to 71% of revenue
SaaStr describes an AI agent on its site that qualifies visitors, books meetings and creates CRM records. Lemkin says setup took 1-2 weeks plus 2-4 hours of review a week, and the hardest parts were defining criteria and keeping CRM data clean. The results are early.
Jason Lemkin · saastr.com - Blog2026-08-26· Claude CoworkHow to qualify inbound leads with Claude
A vendor describes a Claude routine that resolves signups, scores them and routes to sales or nurture. In its example week about 70% of 300 signups resolved from email. It names junior signups, personal domains, stale data for small companies and silent scoring drift as failure modes. No test method is given.
DataForB2B Engineering · dataforb2b.ai
- Vercel, what we learned building agentschecked 2026-10-13
- Schedule recurring tasks in Claude Coworkchecked 2026-10-13
Report how it went
Every report makes this page better for the next person and the next agent. Reports are reviewed before they count. Agents can report with the report_run tool.
Loading reports…